Explainable Boosting Machine (EBM) Model

This notebook shows an example of how to use the APIs with the ExplainableBoostingMachine model from the interpret package and how to get a factual explanation for the model in the form of feature importance.

[1]:

# Load and join raw data sources and their metadata.
%run Example_InputDataSources.ipynb

[2]:

# Joined DataFrame.
df_all.head()

[2]:

	id_c	education_background_c	professional_experience_c	skills_c	gender_c	agg_perceived_foreign_c	id_j	education_reqs_j	experience_reqs_role_j	experience_reqs_duration_j	skills_j	gender_j	agg_perceived_foreign_j	ranking	shortlisted	score
0	5	[{'institution': 'Complutense University Of Ma...	[{'institution': 'Stylo Milano', 'start_date':...	[Communications, Social Integration, Microsoft...	Man	No	5	[Law Bachelor, Degree In Law, Higher Degree In...	[Consultant]	12	[Punctuality, Organization, Accounting, Englis...	Man	No	4	1	0.0
1	6	[{'institution': 'Coronel Rosales Agricultural...	[{'institution': 'Securitas Direct', 'start_da...	[Refinancing, Economy, Microsoft Excel, Collec...	Man	No	3	[]	[Sales Assistant, Saleswoman, Commercial Advisor]	12	[English, Spanish, Communications, Communicati...	Man	No	8	1	0.6
2	10	[{'institution': 'Complutense University Of Ma...	[{'institution': 'Carrefour Express', 'start_d...	[Entrepreneurship, Literacy, Web Design, Adobe...	Woman	No	5	[Law Bachelor, Degree In Law, Higher Degree In...	[Consultant]	12	[Punctuality, Organization, Accounting, Englis...	Man	No	4	1	0.0
3	11	[{'institution': 'Les Ribera De Los Molinos', ...	[{'institution': 'Decimas Sl', 'start_date': '...	[Consulting, Sap Crm, Collections, Automation,...	Woman	No	3	[]	[Sales Assistant, Saleswoman, Commercial Advisor]	12	[English, Spanish, Communications, Communicati...	Man	No	12	0	0.4
4	15	[{'institution': 'Escuela Politcnica Superior ...	[{'institution': 'Reintegrate', 'start_date': ...	[Microsoft Word, Biofuels, English, Entreprene...	Man	No	3	[]	[Sales Assistant, Saleswoman, Commercial Advisor]	12	[English, Spanish, Communications, Communicati...	Man	No	5	1	0.7

[3]:

# Joined metadata.
md_all

[3]:

{'id_c':
        SCHEMA = {'type': 'number'}
        ATTR_TYPE = object
        ATTR_USAGE = default
        KNOWLEDGE_BASE = None,
 'education_background_c':
        SCHEMA = {'type': 'array', 'items': {'type': 'object', 'properties': {'institution': {'type': 'string'}, 'end_date': {'type': 'string'}, 'degree': {'type': 'string'}, 'duration': {'type': 'string'}}}}
        ATTR_TYPE = object
        ATTR_USAGE = default
        KNOWLEDGE_BASE = None,
 'professional_experience_c':
        SCHEMA = {'type': 'array', 'items': {'type': 'object', 'properties': {'institution': {'type': 'string'}, 'end_date': {'type': 'string'}, 'role': {'type': 'string'}, 'duration': {'type': 'string'}}}}
        ATTR_TYPE = object
        ATTR_USAGE = default
        KNOWLEDGE_BASE = None,
 'skills_c':
        SCHEMA = {'type': 'array', 'items': {'type': 'string'}}
        ATTR_TYPE = object
        ATTR_USAGE = default
        KNOWLEDGE_BASE = None,
 'gender_c':
        SCHEMA = {'enum': ['Man', 'Woman', 'Any']}
        ATTR_TYPE = category
        ATTR_USAGE = sensitive
        KNOWLEDGE_BASE = None,
 'agg_perceived_foreign_c':
        SCHEMA = {'enum': ['No', 'Yes', 'Any']}
        ATTR_TYPE = category
        ATTR_USAGE = sensitive
        KNOWLEDGE_BASE = None,
 'id_j':
        SCHEMA = {'type': 'number'}
        ATTR_TYPE = object
        ATTR_USAGE = default
        KNOWLEDGE_BASE = None,
 'education_reqs_j':
        SCHEMA = {'type': 'array', 'items': {'type': 'string'}}
        ATTR_TYPE = object
        ATTR_USAGE = default
        KNOWLEDGE_BASE = None,
 'experience_reqs_role_j':
        SCHEMA = {'type': 'array', 'items': {'type': 'string'}}
        ATTR_TYPE = object
        ATTR_USAGE = default
        KNOWLEDGE_BASE = None,
 'experience_reqs_duration_j':
        SCHEMA = {'type': 'number'}
        ATTR_TYPE = object
        ATTR_USAGE = default
        KNOWLEDGE_BASE = None,
 'skills_j':
        SCHEMA = {'type': 'array', 'items': {'type': 'string'}}
        ATTR_TYPE = object
        ATTR_USAGE = default
        KNOWLEDGE_BASE = None,
 'gender_j':
        SCHEMA = {'enum': ['Man', 'Woman', 'Any']}
        ATTR_TYPE = category
        ATTR_USAGE = sensitive
        KNOWLEDGE_BASE = None,
 'agg_perceived_foreign_j':
        SCHEMA = {'enum': ['No', 'Yes', 'Any']}
        ATTR_TYPE = category
        ATTR_USAGE = sensitive
        KNOWLEDGE_BASE = None,
 'score':
        SCHEMA = {'type': 'number'}
        ATTR_TYPE = numeric
        ATTR_USAGE = target
        KNOWLEDGE_BASE = None,
 'ranking':
        SCHEMA = {'type': 'integer'}
        ATTR_TYPE = ordinal
        ATTR_USAGE = target
        KNOWLEDGE_BASE = None,
 'shortlisted':
        SCHEMA = {'type': 'integer'}
        ATTR_TYPE = category
        ATTR_USAGE = target
        KNOWLEDGE_BASE = None}

[4]:

# Importing libraries to avoid warnings at running time
import warnings
warnings.simplefilter(action='ignore', category=FutureWarning)

[5]:

from sklearn import set_config
set_config(transform_output = "pandas")

[6]:

# Setting category columns in DataFrame based on metadata.
cat_cols = [k for k, v in md_all.items() if v.attr_type=='category']
df_all[cat_cols] = df_all[cat_cols].astype('category')
# Dataframe metadata.
df_all.info()

<class 'pandas.core.frame.DataFrame'>
RangeIndex: 1643 entries, 0 to 1642
Data columns (total 16 columns):
 #   Column                      Non-Null Count  Dtype
---  ------                      --------------  -----
 0   id_c                        1643 non-null   int64
 1   education_background_c      1643 non-null   object
 2   professional_experience_c   1643 non-null   object
 3   skills_c                    1643 non-null   object
 4   gender_c                    1643 non-null   category
 5   agg_perceived_foreign_c     1643 non-null   category
 6   id_j                        1643 non-null   int64
 7   education_reqs_j            1643 non-null   object
 8   experience_reqs_role_j      1643 non-null   object
 9   experience_reqs_duration_j  1643 non-null   int64
 10  skills_j                    1643 non-null   object
 11  gender_j                    1643 non-null   category
 12  agg_perceived_foreign_j     1643 non-null   category
 13  ranking                     1643 non-null   int64
 14  shortlisted                 1643 non-null   category
 15  score                       1643 non-null   float64
dtypes: category(5), float64(1), int64(4), object(6)
memory usage: 149.9+ KB

[7]:

# Define ids, target feature(s), and predictive features.
id_cols = ['id_j', 'id_c']
target_cols = ['score', 'ranking', 'shortlisted']
pred_cols = df_all.columns.difference(target_cols + id_cols)

[8]:

from findhr.preprocess.example_mappings import RelevantExperienceForRole, ExtractMonthDurationJob, MatchOrdinal, \
    ExtractListOfProperty, MatchFeatureAtLeastInList, MatchFeatureSet, MatchBinary

# Calculated features.
maps_derived_1 = {
    (('professional_experience_c', 'experience_reqs_role_j',), ('relevant_exp_role_c',)): RelevantExperienceForRole(),
}

maps_derived_2 = {
    (('relevant_exp_role_c',), ('role_duration_months_c',)): ExtractMonthDurationJob(duration_key='duration_months'),
    (('education_background_c',), ('degree_list_c',)): ExtractListOfProperty(property_key='degree')
}

# Fitness features about the matching between candidate's features and job's requirements.
maps_matching = {
    (('experience_reqs_duration_j', 'role_duration_months_c'), ('fitness_experience',)): MatchOrdinal(),
    (('education_reqs_j', 'education_background_c'), ('fitness_education',)): MatchFeatureAtLeastInList(),
    (('skills_j', 'skills_c'), ('fitness_skills',)): MatchFeatureSet(),
    (('gender_j', 'gender_c'), ('fitness_gender',)): MatchBinary(),
    (('agg_perceived_foreign_j', 'agg_perceived_foreign_c'), ('fitness_foreign',)): MatchBinary()
}

# Helper variable for the fitness features
list_cols_fitness = ['fitness_experience', 'fitness_education', 'fitness_skills', 'fitness_gender', 'fitness_foreign']
maps_matching

[8]:

{(('experience_reqs_duration_j', 'role_duration_months_c'),
  ('fitness_experience',)): MatchOrdinal(),
 (('education_reqs_j', 'education_background_c'),
  ('fitness_education',)): MatchFeatureAtLeastInList(),
 (('skills_j', 'skills_c'), ('fitness_skills',)): MatchFeatureSet(),
 (('gender_j', 'gender_c'), ('fitness_gender',)): MatchBinary(),
 (('agg_perceived_foreign_j', 'agg_perceived_foreign_c'),
  ('fitness_foreign',)): MatchBinary()}

[9]:

# Scikit-learn transformation for numeric and categorical features
from sklearn.pipeline import Pipeline
from sklearn.impute import SimpleImputer
from sklearn.preprocessing import OneHotEncoder, StandardScaler
from sklearn.tree import DecisionTreeRegressor
from sklearn.compose import ColumnTransformer

numeric_features = list_cols_fitness
categorical_features = ['gender_c', 'agg_perceived_foreign_c']
# imputing and scaling numeric features
numeric_transformer = Pipeline(
    steps=[
        ("imputer", SimpleImputer(strategy="median")),  # Not needed for the used dataset.
        ("scaler", StandardScaler())  # Not needed for the decision tree, let's keep it for the sake of generality.
    ]
)
# imputing and encoding categorical features
categorical_transformer = Pipeline(
    steps=[
        ("imputer", SimpleImputer(strategy="most_frequent")),
        # Not needed for the used dataset, again for the sake of generality.
        ("encoder", OneHotEncoder()),  # Convert to one-hot encoding
    ]
)
# combining the two above
column_preprocessor = ColumnTransformer(
    transformers=[
        ("num", numeric_transformer, numeric_features),
        # ("cat", categorical_transformer, categorical_features)
    ],
)

[10]:

from findhr.preprocess.mapping import AttachMetadata, DetachMetadata, DerivedColumn

# The pipeline is composed of two phases:
# 1. Preprocessing with metadata (using findhr package)
pipeline_derived = Pipeline(steps=[
    ("init", AttachMetadata(md_all)),
    ('mapping_1', DerivedColumn(maps_derived_1)),
    ('mapping_2', DerivedColumn(maps_derived_2)),
    ("matching", DerivedColumn(maps_matching)),
    # ("fitness", GroundTruthLinearWeightedScorer(gt_weights_fair)),
    ("end", DetachMetadata())
])
# 2. Standard scikit-learn preprocessing to prepare the data for the model covered by column preprocessor.

[11]:

## Pipeline Including ExplainableBoostingRegressor

[12]:

# Pipeline definition for regression model on the target feature "score".
from findhr.preprocess.mapping import AttachMetadata, DerivedColumn, DetachMetadata
from interpret.glassbox import ExplainableBoostingRegressor
pipeline_regr = Pipeline(
    steps=[
        # first phase: preprocessing with metadata
        ('fitness_value', pipeline_derived
         ),
        # second phase: preprocessing without metadata (standard scikit-learn)
        ("column_preprocessor", column_preprocessor),
        # model inference
        ("regressor", ExplainableBoostingRegressor())
       ]
)

[13]:

# Model fit.
pipeline_regr.fit(df_all.loc[:, pred_cols], df_all.loc[:, 'score'])

[14]:

# Example model prediction.
pipeline_regr.predict(df_all.loc[:10, pred_cols])

[14]:

array([-1.26254049e-17,  6.00000000e-01, -1.23256533e-17,  4.00000000e-01,
        7.00000000e-01, -1.23256533e-17,  3.50000000e-01, -1.26254049e-17,
       -1.23256533e-17,  7.50000000e-01, -1.23256533e-17])

Example Model Explanation

[15]:

### Get a global explanation from ExplainableBoostingRegressor
explanation_global = pipeline_regr.named_steps['regressor'].explain_global()#name=list_cols_fitness)

[16]:

# Visualize the global explanation through plotting the feature importance.
explanation_global.visualize()

[17]:

# Get the transformed data at the end before the model prediction.
idx_explicand_sample = 0
explicand_sample = df_all.loc[:, pred_cols].iloc[idx_explicand_sample:idx_explicand_sample+1]
transformed_data = pipeline_regr[:-1].transform(explicand_sample)
explanation_local = pipeline_regr.named_steps['regressor'].explain_local(transformed_data)

[18]:

# Visualize the local explanation for the first sample explained.
# See documentation at https://interpret.ml/docs/ebm.html for further details.
explanation_local.visualize(0)

[ ]: